Detecting Code Smells using Deep Learning

A smell in software refers to a symptom introduced in software artifacts such as architecture, design or code. A code smell can potentially cause deeper and serious problems, while dealing with mainly non-functional requirements such as testability, maintainability, extensibility and scalability. The detection of code smell is an essential step in the refactoring process, which facilitates non functional requirements in a software. The existing approaches for detecting code smells use detection rules or standards using a combination of different object-oriented metrics. Although a variety of code smell detection tools have been developed, they still have limitations and constraints in their capabilities. The most well-known object-oriented metrics are considered to identify the presence of smells in software. This paper proposes a deep learning based approach to detect two code smells (Brain Class and Brain Method). The proposed system uses thirty open source Java projects, which are shared by many users in GitHub repositories. The dataset of these Java projects is partitioned into mutually exclusive training and test sets. Our experiments have demonstrated high accuracy results for both the code smells.